Newsgroups: comp.ai.neural-nets Subject: RBF NN traning data set limit From: Sergio Ferlito <ferlito.ser...@gmail.com> Date: Wed, 20 Mar 2013 03:59:42 -0700 (PDT) Local: Wed, Mar 20 2013 6:59 am
On Mar 20, 6:59 am, Sergio Ferlito <ferlito.ser...@gmail.com> wrote: > I'm new to RBF Neural Networks and I'm facing the following issue: > in designing a rbf nn for forecasting a variable dependent by two inputs variables , if the traning data set excede 4000 samples, at least using Matlab newrbe function, I'm not able to get a result as Matlab stop responding and also hangs my computer (cpu utilization reaches 90% and memory utilization almost fill all available ram). > I wonder if this is a limitation of RBF NN or a problem to Matlab implementation of RBF NN or ............probably I have committed a design error. > Thanks.
1. Why didn't you crosspost to the MATLAB NEWSGROUP or MATLAB ANSWERS? 2. You might consider NEWRB instead of NEWRBE 3. It is doubtfull that you need 4000 training pairs a. Standardize (zero-mean/unit-variance) x1 and x2 b. Plot y vs x1, y vs x2 and y vs (x1,x2) c. How many points do you really need to adequately represent both I-O relationships? 4. Randomly partition the data into training, validation and test subsets Typically, Ntrn is large enough to adequately represent the plots in 1b and Ntst = Nval. An initial split might be (1334,1333,1333) or (2666,667,667) and, if unsuccessful, changed. 5. Loop a. Create a lot of nets for a full range of Gaussian half-widths. b. Rank the nets using the validation set and choose a "best" net. c. Evaluate the best net using the test set. d. If unsuccessful, choose another random partition and repeat.
P.S. Some people CHEAT: They do not use a test set and choose the best net obtained from the validation set. However, since the validation set is twice as big as before, I'm not sure if the bias is worth worrying about.